Retrieving enhanced item attribute generation

Abstract

Retrieval enhanced item attribute generation techniques are described. The attribute generation system receives an input describing a target item, such as a title of the target item, an image depicting the target item, or a combination thereof. Given the input, the attribute generation system generates a potential spatial embedding representation of the target item and identifies similar items based on the potential spatial embedding representation. The attribute generation system then identifies, for each similar item, one or more aspects that include information describing the similar item. Similar item aspects and inputs describing the target item are used to generate prompts that cause a machine learning system to generate name-value attribute pairs for the target item. A name-value attribute pair is output for display in a user interface and selectable for inclusion in a digital marketplace list for a target item.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to retrieval augmented item attribute generation. BACKGROUND

[0002] Digital marketplaces are increasingly used by entities such as individuals, companies, and the like to facilitate transactions involving items that are listed for sale on the digital marketplace. Accurately describing items for sale with the correct attributes is critical to ensuring trustworthy transactions, as doing so reduces misunderstandings, returns, and complaints. Accurately describing items further improves the visibility and searchability of listings within the digital marketplace, and is therefore desirable for both buyers and sellers participating in digital marketplace transactions. SUMMARY

[0003] Retrieval augmented item attribute generation techniques are described. An attribute generation system receives input describing a target item, such as a title of the target item, an image depicting the target item, or a combination thereof. Given the input, the attribute generation system generates a latent space embedding representation of the target item and identifies similar items based on the latent space embedding representation. The attribute generation system then identifies, for each similar item, one or more aspects that include information describing the similar item. The similar item aspects and the input describing the target item are used to generate a prompt that causes a machine learning system to generate name-value attribute pairs for the target item.

[0004] For example, the attribute generation system generates a large language model (LLM) prompt by updating fields of a template that instructs the LLM to derive name-value attribute pairs for the target item by using the input describing the target item as a primary point of reference to be cross-referenced with similar item aspects as identified by the attribute generation system. The prompt further instructs the LLM to prioritize plausible name-value attribute pairs for the target item in a defined order of importance and assign values to each attribute name-value pair.

[0005] In some implementations, the prompt constrains the LLM to avoid generating attribute names or values that are not included in the similar item aspects and aggregates multiple plausible values for an attribute into a set of candidate values, one of which is automatically selected by the attribute generation system or manually selected via user input. The name-value attribute pairs are output for display in a user interface and are selectable for inclusion in a digital marketplace listing for the target item.

[0006] The summary introduces a series of concepts that will be further described below in the DETAILED DESCRIPTION. As such, the summary is not intended to identify key or essential features of the claimed subject matter, nor is it meant to limit the scope of the claimed subject matter in any way. BRIEF DESCRIPTION OF DRAWINGS

[0007] The specific implementations are described with reference to the accompanying drawings. The entities shown in the figures indicate one or more entities, and thus can be referred to interchangeably in the discussion in singular or plural form.

[0008] Figure 1 is an illustration of a digital media environment in an example implementation that is operable to employ the search-enhanced item attribute generation techniques described herein.

[0009] Figure 2 depicts a system in an example implementation that shows in more detail Figure 1 the operation of the attribute generation system of

[0010] Figure 3 depicts a system in an example implementation that shows in more detail Figure 2 the operation of the prompt generation module of Figure 1 based on aspects of one or more items identified by the attribute generation system of

[0011] Figure 4 depicts a system in an example implementation that shows Figure 1 the output of the user interface of Figure 1 based on a prompt generated by the attribute generation system of

[0012] Figure 5 depicts a system in an example implementation that shows the operation of the list generation system to generate an item list for a target item that includes item attributes generated by the attribute generation system of Figure 1

[0013] Figure 6 depicts a process in an example implementation of the attribute generation system of Figure 1 to generate item attributes that describe a target item based on input that identifies the target item.

[0014] Figure 7 shows an example system that includes various components of an example device that can be implemented as any type of computing device described with reference to the previous figures and / or utilized to implement the techniques described herein. DETAILED DESCRIPTION

[0015] SUMMARY

[0016] ​Accurately describing items for sale with the correct attributes is critical to facilitating trustworthy transactions on digital marketplaces because doing so reduces misunderstandings, returns, and complaints. Accurately describing items further improves the visibility and searchability of listings within digital marketplaces and is therefore desirable for both buyers and sellers participating in digital marketplace transactions.

[0017] However, conventional techniques for generating item listings rely on the knowledge of the seller and force the seller to manually enter detailed product information when defining aspects of the item being listed for sale. As a result, these conventional techniques often involve searching for product information that the seller does not know or does not readily have access to, which is particularly time consuming for sellers listing multiple items for sale. Moreover, given the variations between different versions of a single product, the seller often does not know which aspects correctly identify the item being listed for sale, which makes the item listing process error prone and forces the seller to conduct additional research, thereby increasing the burden on the seller to list items and decreasing market participation.

[0018] To address these technical challenges faced by conventional systems and techniques, enhanced item attribute retrieval generation techniques are described. For example, an attribute generation system is configured to implement an item embedding module, a similarity module, an aspect retrieval module, a prompt generation module, and an aspect generation module that collectively generate item attributes based on inputs describing a target item, such as a title of the target item, an image describing the target item, or a combination thereof.

[0019] Based on the inputs describing the target item, the item embedding module generates a latent space embedding representation of the target item, which is used by the similarity module to identify similar items to the target item. The list of similar items generated by the similarity module is used by the aspect retrieval module to generate one or more aspects for each item in the list of similar items, where the aspects include information describing the respective similar item.

[0020] The prompt generation module then generates a prompt using the inputs describing the target item and the aspects of the similar items. In implementations, generating the prompt involves updating fields of a template designed to cause one or more machine learning models (e.g., a large language model (LLM)) to generate an output that includes name-value pairs that accurately characterize attributes of the target item. The name-value attribute pairs generated by the one or more machine learning models are then presented to be output in a user interface (e.g., at a display device of a computing system used by the seller to generate the item listing).

[0021] In some implementations, the attribute generation system is supported by a service provider system that further supports a listing generation system. In such implementations, the service provider system is configured to support user interaction in selecting name-value attribute pairs for inclusion in an item listing that lists a target item for sale on one or more digital marketplaces. For example, a user interface that displays name-value attribute pairs generated by the attribute generation system facilitates selection of one or more of the name-value pairs, and displays an item listing for the target item that includes the selected attribute name-value pairs.

[0022] The item listing for the target item with the selected attribute name-value pairs is then published via one or more digital marketplaces. In this way, the retrieval-enhanced item attribute generation techniques described herein address technical challenges associated with conventional item listing techniques, thereby providing improved user and computational efficiency. Further discussion of these and other examples is included in the following description and illustrated in relation to the accompanying drawings.

[0023] In the following discussion, an example environment is described that is configured to employ the techniques described herein. Example processes that are configured for execution in the example environment, as well as other environments, are also described. Thus, the execution of the example processes is not limited to the example environment and the example environment is not limited to the execution of the example processes.

[0024] Example Environment

[0025] Figure 1 FIG. 1 is an illustration of a digital media environment 100 in an example implementation that is operable to employ the retrieval-enhanced item attribute generation techniques described herein. The illustrated environment 100 includes a service provider system 102 and a computing device 104 communicatively coupled to each other via a network 106. The computing device, for example, one or more computing devices represented by the service provider system 102 and / or the computing device 104, can be configured in a variety of ways.

[0026] For example, a computing device can be configured as a desktop computer, a laptop computer, a mobile device (e.g., assuming a hand-held configuration such as a tablet computer or a mobile phone), etc. Thus, a computing device ranges from full resource devices with substantial memory and processor resources (e.g., personal computers, game consoles) to a low-resource device with limited memory and / or processing resources (e.g., mobile devices). Additionally, although a single computing device is shown and described in instances of the following discussion, a computing device also represents a plurality of different devices, for example, utilized by an entity to perform "on the cloud" operations for the service provider system 102 and as described in further detail with respect to Figure 7 The service provider system 102 is further described in further detail below.

[0027] The service provider system 102 represents a combination of hardware and software resources (e.g., instructions stored on a computer-readable storage medium that are executed by at least one processing device) to provide a digital service (e.g., a digital service that is remotely available to the computing device 104 via the network 106). As an example, the digital service provided by the service provider system 102 includes a digital marketplace platform, such as a cloud-based, modular architecture that enables secure and scalable transactions between buyers and sellers that list products or services for sale via a digital marketplace listing.

[0028] In implementations, such a digital service provided by the service provider system 102 is built on microservices that handle various digital marketplace functions, such as user authentication, item listing management, payment processing, order fulfillment, and the like. In some implementations, the digital service provided by the service provider system 102 is scalable to third-party integrations such that functionality of the service provider system 102 is integrated or otherwise extended to other digital services. As described below with reference to FIG. 2, the service provider system 102 includes a set of microservices that are configured to provide a digital marketplace service. Figure 7 In more detail, the service provider system 102 represents a distributed data storage system that is configured to ensure fault tolerance and implement encryption protocols to protect sensitive user data, transaction data, and the like.

[0029] In the specific example context of a digital marketplace service provided by the service provider system 102, the service provider system 102 incorporates search and recommendation algorithms to personalize the experience of users interfacing with the digital marketplace of the service provider system 102. Continuing with this example context of a digital marketplace service, the service provider system 102 implements an access control system to ensure that certain resources are limited to authorized entities, provide analytics for marketplace activity, and the like.

[0030] In Figure 1 In the illustrated example, the computing device 104 includes a communication module 108 for accessing a digital service provided by the service provider system 102 (e.g., via the network 106). For example, the communication module 108 represents a browser configured to access a digital marketplace of the service provider system 102 via the Internet, an application provided by the service provider system 102, a combination thereof, and the like. Thus, the communication module 108 represents functionality of the computing device 104 to transmit data to and receive data from the service provider system 102.

[0031] The service provider system 102 is depicted as including an attribute generation system 110. Although illustrated and described herein as being implemented at the service provider system 102, in some implementations, the attribute generation system 110 is implemented locally at the computing device 104. The attribute generation system 110 implements a machine learning system 112 and is configured to receive an input 114 describing a target item to be listed for sale via a digital marketplace. Based on the input 114, the attribute generation system 110 generates an output including item attributes 116 of the target item using the machine learning system 112 (e.g., the attribute generation system 110 outputs item attributes of the target item using one or more trained generative artificial intelligence models). The item attributes 116 represent name-value attribute pairs describing characteristics of the target item, where the described characteristics of the target item are not included in the input 114 describing the target item.

[0032] The computing device 104 is also configured to include a display device 118. The display device 118 represents hardware configured to output visual information to one or more users of the computing device 104. For example, the display device 118 represents a monitor (e.g., an external screen connected to the computing device 104 and optionally integrated into the form factor of the computing device 104). Alternatively or additionally, the display device 118 represents a touchscreen for outputting visual information and receiving input from a user of the computing device 104. Alternatively or additionally, the display device 118 represents a projector configured to display visual information on a relatively large surface. Alternatively or additionally, the display device 118 represents a wearable device (e.g., a virtual reality headset) that immerses a user of the computing device 104 in a digital environment. Other examples are contemplated according to the described technology such that the display device 118 represents a range of different display sizes, resolutions, and configurations for outputting visual information on behalf of the computing device 104.

[0033] For example, the display device 118 outputs a display of a user interface 120 for the service provider system 102. In the illustrated example, the user interface 120 is depicted as displaying visual information describing the input 114 provided to the service provider system 102 as well as the item attributes 116 generated from the input 114. For example, the user interface 120 includes a target item image 122 as well as a target item title 124. In the illustrated example, the user interface 120 also includes a target item description 126. The target item description 126 represents textual information describing the target item. For example, the target item description 126 includes a target item description text 128. In the illustrated example, the target item description 126 also includes a target item description image 130. In the illustrated example, the target item description 126 also includes a target item description video 132. Figure 1 Figure 1 ​In the illustrated example, the target item for which an item listing is to be generated is a phone, with the target item image 122 depicting the phone and the target item title 124 providing a textual description of the phone. In comparison to the specific information included in the item listing for the target item, the target item title 124 describes the phone in high-level summary (e.g., “regular version 11 phone”). In implementations, the input 114 is provided by a user of the computing device 104, such as by a seller seeking to list the target item for sale via a digital marketplace provided by the service provider system 102.

[0034] For example, the target item image 122 represents a photograph of the phone taken by the user of the computing device 104 (e.g., using an image capture device, such as a camera, integrated into the computing device 104), and the target item title 124 represents a textual description of the phone provided by the user of the computing device 104. In implementations, the input 114 includes one or more of the target item image 122 or the target item title 124 (e.g., the input 114 includes only an image of the phone, only a general textual description of the phone, or both). In this way, the input 114 represents information that describes the target item in high-level summary, which lacks one or more specific aspects that distinguish the target item from other items (e.g., aspects that distinguish the specific phone depicted in the target item image 122 from other phones).

[0035] As described in greater detail below, the attribute generation system 110 utilizes the machine learning system 112 to generate item attributes 116 based on the input 114. The item attributes 116 are communicated from the service provider system 102 to the computing device 104 (e.g., for display in the user interface 120). As described in greater detail below, the item attributes 116 are generated by the attribute generation system 110 based on the input 114, which represents information that describes the target item in high-level summary, which lacks one or more specific aspects that distinguish the target item from other items. Figure 1 In the illustrated example, the item attributes 116 generated by the attribute generation system 110 are returned as name-value attribute pairs 126. In the context of “name-value attribute pairs,” the term “name” refers to a label or identifier that specifies a particular attribute of an item, such as “color” or “condition.” Thus, the “name” of a name-value attribute pair defines a characteristic of the target item being described. The term “value” represents specific information corresponding to the named attribute, such as “blue” for color or “used” for condition. The name and value together function as a pair, with the name indicating the attribute and the value quantifying and / or qualifying the data associated with the indicated attribute, thereby providing a structured form for describing various attributes of an item.

[0036] As a specific example, Figure 1The item attributes 116 depicting the phone represented by the input 114 include the following name-value attribute pairs 126: "Condition: New"; "Processor: Hypercore"; "Screen Size: 7 inches"; "Feature: Accelerometer"; "Feature: Barometer"; "Feature: Bluetooth"; "Feature: Facial Recognition"; "Feature: Proximity Sensor"; and "Feature: Wireless Charging". Thus, the item attributes 116 represent information that describes the target item in additional detail beyond the general description of the target item represented by the input 114. In implementations, each of the attributes in the item attributes 116 is selectable (e.g., via user input at the computing device 104) to be included in a listing of the target item (e.g., the phone depicted by the target item image 122) for sale via a digital marketplace of the service provider system 102. In contrast to conventional item listing techniques that require a seller to manually curate and input item attributes to be included in an item listing, the attribute generation system 110 automatically generates the item attributes 116 that accurately describe the target item in specific detail. For further description of the attribute generation system 110 that generates the item attributes 116, consider Figure 2 .

[0037] Generally, the functions, features and concepts described with respect to the above and below examples are employed in the context of the example processes described in this section. Moreover, the functions, features and concepts described in this document with respect to different figures and examples can be interchanged and are not limited to implementation in the context of a particular figure or process. Furthermore, the blocks associated with different representative processes and corresponding figures herein can be applied together and / or combined in different ways. Thus, the various functions, features and concepts described herein with respect to different example environments, devices, components, figures and processes can be used in any suitable combination and are not limited to the particular combinations represented by the examples listed in this specification.

[0038] Example Retrieval Enhanced Item Property Generation

[0039] Figure 2 The system 200 in the example implementation is depicted in greater detail showing operations of the attribute generation system 110 generating item attributes 116 that describe a target item based on an input 114 that identifies the target item. In the example shown, the attribute generation system 110 is depicted as including an item embedding module 202. The item embedding module 202 represents functionality of the attribute generation system 110 to generate an embedding 204 for a target item described by the input 114. The embedding 204 represents a latent space embedding representation of the target item. Figure 2

[0040] ​For example, in one or more implementations, the item embedding module 202 represents a service built using a representational state transfer (REST) architectural style. Thus, the item embedding module 202 represents a RESTful service supported by the attribute generation system 110 that leverages one or more deep learning models to generate an embedding 204 based on the input 114, where the embedding 204 outputs a dense vector representation of a target item described by the target item image 122, the target item title 124, or a combination thereof.

[0041] Given the input 114, the item embedding module 202 processes the target item image 122, the target item title 124, or a combination thereof by leveraging a deep learning model configured with an architecture such as a transformer (for text), a convolutional neural network (for images), or a combination thereof. In this way, the embedding 204 represents a fixed-dimension array of floating-point numbers that captures semantic attributes of the target item as described by the input 114.

[0042] In implementations, the item embedding module 202 is trained using a database that pairs items with their known titles and / or images such that each item in the database is associated with a ground truth label. This labeled data (e.g., by the attribute generation system 110) is used to train a multi-modal model that includes a text encoder and an image encoder such that the item embedding module 202 is configured to process both the target item image 122 and the target item title 124. For example, the text encoder represents a function of the item embedding module 202 to transform the target item title 124 into an embedding (e.g., using a bidirectional encoder representations from transformers (BERT) architecture from a transformer). The image encoder represents a function of the item embedding module 202 to transform the target item image 122 into an embedding (e.g., using a concurrent neural network architecture such as ResNet or EfficientNet).

[0043] The text and image embeddings generated by the item embedding module 202 are projected into a common latent space. In implementations, during training of one or more machine learning models implemented by the item embedding module 202, a contrastive loss (e.g., a triplet loss) is employed to ensure that items with similar titles and images have embeddings that are close in the latent space (e.g., a vector space) while dissimilar items are mapped further apart from each other. The image encoder and the text encoder of the item embedding module 202 are updated via backpropagation using the labeled dataset that ensures the embeddings remain semantically relevant. In implementations, the attribute generation system 110 further refines the item embedding module 202 using augmentations such as data shuffling, dropout, and regularization to improve the embeddings 204 output by the item embedding module 202.

[0044] The attribute generation system 110 is also depicted as including a similarity module 206. The similarity module 206 represents functionality of the attribute generation system 110 to generate a list of similar items 208 based on the embedding 204, where the list of similar items 208 includes a plurality of items that are similar to the target item identified by the input 114. In implementations, the similarity module 206 is configured to maintain a RESTful service of an Approximate Nearest Neighbor (ANN) index. The ANN index maintained by the similarity module 206 is designed to efficiently retrieve similar items based on their respective embeddings (e.g., retrieve items having embeddings similar to the embedding 204). The functionality of the similarity module 206 involves inputting the embedding 204 to a REST endpoint and querying the ANN index to identify nearest items (e.g., identify similar items based on cosine similarity).

[0045] In implementations, the similarity module 206 utilizes one or more ANN algorithms to generate the list of similar items 208, such as a Hierarchical Navigable Small World (HNSW) or Locality- sensitive Hashing (LSH) algorithm. By utilizing one or more ANN algorithms that use the ANN index, the similarity module 206 efficiently identifies items that are similar to the target item identified by the input 114. Advantageously, the similarity module 206 generates the list of similar items 208 with sublinear time complexity relative to conventional search methods, such as brute-force methods, thereby making the functionality of the similarity module 206 scalable for large ANN indexes.

[0046] In implementations, to generate the list of similar items 208, the similarity module 206 normalizes the vector representation of the embedding 204 to ensure consistent cosine similarity calculations. Cosine similarity, which measures the angle between two vectors in the embedding space, is used to quantify the similarity between the embedding 204 and respective embeddings of items included in the list of similar items 208. For example, the cosine distance (1 - cosine similarity) is used, and the goal of the similarity search performed by the similarity module 206 is to minimize this distance.

[0047] The result of the similarity search performed by the similarity module 206 is returned as a list of item identifiers that represent the nearest neighbors to the embedding 204. In some implementations, the attribute generation system 110 delegates the similarity module 206 to sort the item identifiers in ascending order of cosine distance, which means that the first item in the list has the smallest distance (i.e., the highest similarity), and each subsequent item has a progressively larger distance from the embedding 204. In implementations, the list of similar items 208 is output in JavaScript Object Notation (JSON) format, containing an array of the top k nearest item identifiers, and their corresponding cosine distances in case additional knowledge of the similarity ranking is needed.

[0048] In implementations, the attribute generation system 110 is configured to update the ANN index used by the similarity module 206 to generate the list of similar items 208, e.g., to incorporate new embeddings as items are added or deleted in the service provider system 102, to ensure that the results reflect the most up-to-date representation of the dataset.

[0049] The attribute generation system 110 also includes an aspect retrieval module 210. The aspect retrieval module 210 represents functionality of the attribute generation system 110 to compare each item identifier included in the list of similar items 208 to entries in an item database 212. The entries in the item database 212 represent groupings of known item identifiers for items and corresponding metadata (e.g., titles, category identifiers, item attributes, prices, etc.) that describe the items.

[0050] In some implementations, the item database 212 represents a local database (e.g., the item database 212 of the service provider system 102) that is accessible to the attribute generation system 110. Alternatively or additionally, the item database 212 is a database associated with one or more different service providers that is accessed by the attribute generation system 110 (e.g., via the network 106). For each item in the list of similar items 208, the aspect retrieval module 210 is configured to output similar item aspects 214, where the aspects include information that describes the respective similar item obtained from the item database 212.

[0051] The similar item aspects 214 and the input 114 are then provided as inputs to a prompt generation module 216 to generate a prompt 218. The prompt 218 is configured to initiate generation of the item attribute 116 based on the input 114 and the similar item aspects 214 using one or more machine learning models of the machine learning system 112. The prompt generation module 216 is configured to implement this by filling out one or more templates 220 (shown as stored in a storage device 222) using natural language processing by the prompt generation module 216. Specific examples of the prompt generation module 216 generating the prompt 218 by filling out the one or more templates 220 are described in more detail below. Figure 3

[0052] The prompt 218 is then provided as input to an aspect generation module 224, which represents functionality of the attribute generation system 110 to output the item attribute 116 (e.g., for display in the user interface 120) using one or more trained machine learning models of the machine learning system 112 (e.g., at least one LLM, as described in more detail below with reference to Figure 4

[0053] Figure 3 ​​The system 300 in an example implementation is depicted, which shows in more detail the operation of the prompt generation module 216 generating the prompt 218 based on the similar item aspects 214 generated by the attribute generation system 110. In the depicted example, the prompt 218 is generated based on two of the one or more templates 220: template 302 and template 304. In the depicted example, the text of the prompt 218 uses parentheses to distinguish the text of the templates 220 from the text of the input 114 or the text of the similar item aspects 214. For example, the prompt generation module 216 generates the prompt 218 by inserting the text of the input 114 and the text of the similar item aspects 214 at the locations of the prompt 218 that are bracketed by the parentheses. Figure 3 In the depicted example, the prompt 218 is generated based on two of the one or more templates 220: template 302 and template 304. In the depicted example, the text of the prompt 218 uses parentheses to distinguish the text of the templates 220 from the text of the input 114 or the text of the similar item aspects 214. For example, the prompt generation module 216 generates the prompt 218 by inserting the text of the input 114 and the text of the similar item aspects 214 at the locations of the prompt 218 that are bracketed by the parentheses.

[0054] For example, the prompt 218 is depicted as qualifying the aspect generation module 224 to define the following purpose: derive attribute name-value pairs for the [target item] by analyzing [target item] as the primary point of reference to cross-reference with information describing [similar items]. In this prompt 218, [target item] is obtained from at least one of the target item image 122 or the target item title 124 represented by the input 114. In some implementations, [target item] is different from the input 114 (e.g., these portions of the prompt 218 are replaced with more accurate descriptions identified by the embedding 204 used by the attribute generation system 110).

[0055] The prompt 218 is further generated to qualify the aspect generation module 224 to define the purpose of instructing the aspect generation module 224 to identify, derive, and prioritize, in order of importance, reasonable attributes for the [target item] by assigning values to each of the attributes. The prompt 218 also ensures that only relevant similar item aspects 214 are included in the item attributes 116 by instructing the aspect generation module 224 to ignore irrelevant items of [similar items] (e.g., irrelevant items that can be inadvertently included in the similar item list 208).

[0056] The hint 218 is further refined using template 304, causing the aspect generation module 224 to ensure that the attribute corresponds to the [target item] by excluding irrelevant reference points, reiterating the importance of ignoring the similar item aspect 214 for items that are not similar to the target item identified by input 114. Hint 218 also instructs the aspect generation module 224 to prioritize attributes existing in the [similar item] reference points by generating new attributes only when necessary and when the new attribute does not exist in the [similar item] reference points. Hint 218 causes the aspect generation module 224 to aggregate multiple reasonable values ​​for the attribute into a list of candidate values. In this implementation, this causes the aspect generation module 224 to output multiple possible values ​​for the name-value attribute pair (e.g., for display in user interface 120), and (e.g., to user input in user interface 120) a hint that the appropriate possible value from the possible values ​​is selected to be included as the attribute name in characterizing item attribute 116.

[0057] Hint 218 further instructs aspect generation module 224 to assign higher importance (e.g., greater mathematical weight) to basic item attributes with physical meaning, such as condition, brand, human observable characteristics, etc. Hint 218 favors attributes lacking specific importance (e.g., those perceived by a person considering purchasing the item, such as Manufacturer Part Number (MPN), Universal Product Code (UPC), etc.). Additionally, hint 218 instructs aspect generation module 224 to present item attribute 116 in a [qualified format] without additional information. In this implementation, the [qualified format] of item attribute 116 (e.g., JSON) is specified by attribute generation system 110, specified by user of computing device 104, or a combination thereof.

[0058] Figure 4 The example implementation of system 400 is depicted, showing the output of user interface 120 as an example of displaying item attributes 116 generated by machine learning system 112 based on prompts generated by attribute generation system 110. Figure 4 In the example shown, aspect generation module 224 is depicted as generating item attributes 116 based on input prompts 218 using LLM 402 of machine learning system 112. LLM 402 represents at least one LLM built on a transformer architecture, such that LLM 402 is designed to handle sequential data and natural language processing tasks.

[0059] In some implementations, the LLM 402 includes a multi-layer self-attention mechanism, where each layer contains two main components: a multi-headed self-attention mechanism and a feed-forward neural network. The self-attention mechanism enables the LLM 402 to attend to different parts of the input sequence (e.g., the prompt 218) simultaneously, capturing dependencies between words or tokens regardless of their position in the sequence. The LLM architecture is scaled to a large number (e.g., billions or even trillions) of parameters, with layers stacked deep (e.g., hundreds of layers) to capture complex patterns and representations.

[0060] In implementations, the LLM implemented by the aspect generation module 224 is pre-trained on a large amount of text data, where the LLM learns to predict missing tokens or next tokens based on context, resulting in an emergence of a rich latent space representation of language. After pre-training, fine-tuning on task- or domain-specific data is utilized to enhance performance for specific applications (e.g., outputting item attributes 116 based on the prompt 218). In implementations, the parameters of the LLM 402 are optimized using variants of stochastic gradient descent (e.g., Adam), enabling the LLM 402 to handle a wide range of natural language understanding and generation tasks.

[0061] As displayed in the user interface 120, the output generated by the LLM 402 indicates that the target item described by the input 114 can be inferred to have the attribute name-value pairs included in the item attributes 116. In some implementations, the output of the LLM 402 also includes information describing the generation of the item attributes 116 based on cross-referencing the input 114 (e.g., the target item image 122, the target item title 124, or a combination thereof) with multiple reference items.

[0062] In some implementations, the LLM 402 is configured to output information describing the multiple reference items from which the item attributes 116 were derived and / or generated. For example, in the user interface 120, the LLM 402 outputs information describing the multiple reference items from which the item attributes 116 were derived and / or generated. Figure 4In the illustrated example of FIG. 4, the user interface 120 includes a control 404 that is selectable (e.g., via user input) to cause display of additional information about one or more particular reference items used to generate the item attributes 116. As a specific example, in response to detecting input selecting the control 404, the attribute generation system 110 outputs (e.g., via the user interface 120) a display of information describing one or more of the items included in the similar items list 208, one or more of the similar items aspects 214 for at least one similar item, or a combination thereof. In this way, the attribute generation system 110 is configured to provide a user (e.g., of the computing device 104) with information describing the basis on which the LLM 402 outputs the item attributes 116. Advantageously, this explanation of the basis for generating the item attributes 116 enhances the transparency and trustworthiness of the attribute generation system 110, allowing the user of the computing device 104 to better understand the reasoning and data sources supporting the generation of the item attributes 116. This output generated by the machine learning system 112 of the aspect generation module 224 further enables the user to assess the accuracy and relevance of the item attributes 116, allowing for a more informed decision-making process when listing a target item for sale.

[0063] Figure 4 The user interface 120 depicted in the example of FIG. 4 also includes example item attributes 116 generated for the target item identified by the input 114. Continuing with the example of the target item being a phone, the item attributes 116 depicted in FIG. 4 include a description of the target item as being a “Renphone v11-512GB-Factory Unlocked” phone. Figure 1 In the illustrative example of FIG. 4, Figure 4 The item attributes 116 depicted in FIG. 4 represent a curated and prioritized list of attributes that specifically describe the phone depicted in the target item image 122 and textually described by the target item title 124.

[0064] Specifically, Figure 4 The item attributes 116 depicted in FIG. 4 specify that the title of the target item for sale to be included in the listing is “Renphone v11-512GB-Factory Unlocked.” The item attributes 116 also describe that the condition of the target item is “New,” that the target item has a “Hypercore” processor, a screen size of “7 inches,” and is associated with the brand “Roger’s RenPhone.” As indicated by the hints 218, the LLM 402 prioritizes the order of these item attributes 116 to favor attributes that are more easily distinguishable or distinguishable from other items (e.g., other items listed for sale via a digital marketplace). Figure 3

[0065] Figure 4 ​The depicted item attributes 116 also include various features of the target item, specifically describing that the phone supports 4K video recording, includes an accelerometer, includes a barometer, supports Bluetooth connectivity, includes hardware and software for performing facial recognition, includes GPS capability, includes a proximity sensor, is configured for wireless charging, etc. Thus, the item attributes 116 generated by the attribute generation system 110 represent information that is different from and much more specific than the highly generalized information included in the input 114 describing the target item. The item attributes 116 automatically generated by the attribute generation system 110 enable the generation of a listing for the target item for sale in a manner that avoids the tedious and error-prone manual steps required by conventional systems.

[0066] Figure 5 The system 500 in the depicted example implementation is shown illustrating the operation of a listing generation system generating an item listing for a target item that includes the item attributes 116 generated by the attribute generation system 110. In the depicted example, the service provider system 102 is depicted as including a listing generation system 502. The listing generation system 502 represents the functionality of the service provider system 102 to generate and publish one or more listings for sale (e.g., of goods and / or services) via one or more digital marketplaces. Figure 5

[0067] For example, the listing generation system 502 represents a service provided by the service provider system 102 that automates one or more steps involved in generating the item listing 504, such as to ensure consistency in the presentation of various item listings at a digital marketplace. The listing generation system 502 is configured to interface with the backend architecture of at least one digital marketplace and utilize a structured data model to organize information describing the subject matter of the listing (e.g., the input 114 provided to the attribute generation system 110, the item attributes 116 generated by the machine learning system 112 of the attribute generation system 110, or a combination thereof). The listing generation system 502 is configured to process the item attributes 116 through a series of validation, formatting, and categorization modules that enforce compliance with one or more standards specified by the digital marketplace. In doing so, the item listing 504 generated by the listing generation system 502 adheres to the aesthetic and structural guidelines of the digital marketplace, such as being consistent with the interface design requirements of the marketplace.

[0068] ​For example, list generation system 502 truncates text associated with one or more of the item attributes 116 to specified character limits, resizes and optimizes images (e.g., target item image 122) for display, and standardizes the organization or placement of item attributes 116 to a predefined format. In some implementations, list generation system 502 combines semantic markup and search engine optimization (SEO) modules to automatically generate metadata tags and optimize item list 504 to achieve search engine discoverability both within digital marketplaces and across external search platforms.

[0069] In its implementation, the list generation system 502 is designed to have at least one layer (e.g., an API layer) that enables the item list 504 to integrate with external inventory management or content management systems (CMS), thereby extending the utility of the item list 504 beyond the single digital marketplace hosting it. Additionally, the list generation system 502 is configured (e.g., via user interface 120) to output a display of the item list 504, allowing the user of the computing device 104 to modify or update the information included in the item list 504 before publishing it in one or more digital marketplaces.

[0070] Example systems and techniques for generating item attributes of target items using retrieval-enhanced generation have been considered; now, example processes are considered to illustrate aspects of the techniques described herein.

[0071] Example Process

[0072] The following discussion describes techniques configured to be implemented using the systems and devices described herein. Each aspect of the process is configured to be implemented in hardware, firmware, software, or a combination thereof. These processes are shown as a set of boxes specifying operations performed by one or more devices, and are not limited to the shown order of operations for performing the individual boxes. References are made in the following sections of the discussion. Figures 1-5 .

[0073] Figure 6 The process 600 of an example implementation of an attribute generation system 110 that generates item attributes describing a target item based on an input identifying the target item is described.

[0074] First, information describing the items to be listed for sale is received (box 602). The attribute generation system 110 receives, for example, input 114 from the computing device 104, including at least one of a target item image 122 or a target item title 124. Then, multiple similar items are identified based on the received information (box 604). For example, the item embedding module 202 generates an embedding 204 based on the input 114. The similarity module 206 then processes the embedding 204 to generate a list of similar items 208.

[0075] For each of the plurality of similar items, at least one aspect describing a characteristic of the similar item is identified (block 606). For example, the aspect retrieval module 210 generates similar item aspects 214 for each item included in the similar item list 208 based on information maintained in the item database 212.

[0076] Then, a prompt is generated to initiate a description of the item to be listed for sale with a different attribute than the received information based on the identified at least one attribute for each of the plurality of similar items (block 608). For example, the prompt generation module 216 generates the prompt 218 using one or more templates 220, such as the template 302 and the template 304.

[0077] The prompt is input to one or more machine learning models to produce an output of an attribute (block 610). For example, the aspect generation module 224 inputs the prompt 218 to the LLM 402 of the machine learning system 112, which causes the machine learning system 112 to output the item attribute 116. The attribute is then presented in a user interface (block 612). For example, the attribute generation system 110 communicates the item attribute 116 to the computing device 104 for presentation in the user interface 120 by the display device 118.

[0078] Having described example processes in accordance with one or more implementations, consider now an example system and device that implements the various techniques described herein.

[0079] Example System and Device

[0080] Figure 7 An example system including an example computing device 702 that is representative of one or more computing systems and / or devices that can implement the various techniques described herein is illustrated at 700. This is illustrated by including the service provider system 102 and the attribute generation system 110. The computing device 702 can be configured, for example, as a server of a service provider, a device associated with a client (e.g., a client device), a system on a chip, and / or any other suitable computing device or computing system.

[0081] The example computing device 702 as illustrated includes a processing device 704, one or more computer-readable media 706, and one or more I / O interfaces 708 that are communicatively coupled with one another. Although not illustrated, the computing device 702 also includes a system bus or other data and command transfer system that couples the various components with one another. A system bus can include any one or combination of different bus structures, such as a memory bus or memory controller, a peripheral bus, a serial bus, and / or a local bus, using any of a variety of bus architectures. Various other examples are also contemplated, such as control lines, data lines, and the like.

[0082] The processing device 704 is representative of the functionality performed by hardware as processing devices 704 execute instructions. As such, the processing device 704 is illustrated as including hardware elements 710 that are configured to perform a variety of functions as described below. One or more of these hardware elements 710 can be configured to act as one or more processors, functional blocks, and so forth. This includes the use of one or more integrated circuits, for example, ASICs. Accordingly, the hardware elements 710 are not limited by the materials from which they are formed or the processing mechanisms employed therein. For example, a processor can be configured to be any of a general purpose microprocessor, any member of the Pentium® family of microprocessors, microcontrollers, embedded microprocessors, RISC processors, CISC processors, and so on. In such a context, the processor can execute instructions as a result of its processing that, for example, implement methods, techniques, functions, procedures, etc.

[0083] The computer-readable storage medium 706 is illustrated as including memory / storage 712 that stores instructions that can be executed by the processing device 704. The computer-readable storage medium 706 is configured for storing instructions that are executable by the processing device 704 to cause the processing device 704 to perform operations. The memory / storage 712 is representative of the memory / storage capacity associated with one or more computer-readable media. The memory / storage 712 includes volatile media (such as random access memory (RAM)) and / or nonvolatile media (such as read only memory (ROM), floppy disks, etc.). The memory / storage 712 includes fixed and removable media, for example. The computer-readable medium 706 can be configured in a variety of other ways as further described below.

[0084] The input / output interface 708 is representative of functionality to enable a user to enter commands and information to computing device 702, and further allows for information to be presented to the user and / or other components or devices using various input / output devices. Examples of input devices include a keyboard, cursor control device (e.g., a mouse), microphone, scanner, touch functionality (e.g., capacitive or other sensors that are configured to detect physical touch), a video capture device (e.g., a camera that is configured to capture still images or video), and so forth. Examples of output devices include a display device (e.g., a monitor or projector), speakers, a printer, a network card, tactile-response device, and so on. Thus, the computing device 702 can be configured in a variety of ways as further described below to support user interaction.

[0085] Various techniques are described herein in the general context of software, hardware elements, or program modules. Generally, such modules include routines, programs, objects, elements, components, data structures, and the like that perform particular tasks or implement particular abstract data types. The terms "module," "functionality," and "component" as used herein generally represent software, firmware, hardware, or a combination thereof. The features of the techniques described herein are platform-independent, meaning that they can be implemented on a variety of commercial computing platforms having a variety of processors.

[0086] Implementations of the described modules and techniques are stored on or transmitted across some form of computer-readable media. The computer-readable media include various media that are accessible by the computing device 702. By way of example, and not limitation, computer-readable media include "computer-readable storage media" and "computer-readable signal media."

[0087] "Computer-readable storage media" refers to media and / or devices that enable persistent and / or non-transitory storage of information (e.g., storage media that store software, data, or other computer-readable instructions in a persistent manner all). Thus, computer-readable storage media refers to non-signal bearing media. Computer-readable storage media includes hardware such as volatile and non-volatile, removable and non-removable media implemented in a method or technology suitable for storage of information such as computer readable instructions, data structures, program modules, logic elements / circuits, or other data. Examples of computer-readable storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, hard disks, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or other storage devices, tangible media, or articles of manufacture suitable to store desired information and which can be accessed by a computer.

[0088] "Computer-readable signal media" refers to a signal-bearing medium that is configured to transmit instructions to the hardware of the computing device 702, such as via network. Signal media typically is embodied in waves, such as those used to carry data

[0089] As previously described, hardware elements 710 and computer-readable media 706 are representative of modules, programmable device logic and / or fixed device logic implemented in hardware that are used to implement at least portions of the techniques described herein, such as for example executing one or more instructions. Hardware can include

[0090] The foregoing combination of software, hardware or executable modules are implemented to realize the various techniques described herein. Thus, software, hardware or executable modules are implemented as one or more instructions and / or logic embodied on some form of computer-readable storage media and / or by one or more hardware elements 710. The computing device 702 is configured to implement particular instructions and / or functions corresponding to the software and / or hardware modules. Implementation of the

[0091] The technology described herein can be supported by various forms of the computing device 702 and is not limited to the specific examples of the technology described herein. The functionality of the technology can be implemented all or in part through use of distributors such as via the platform 716 by the "cloud" 714, as described below.

[0092] The cloud 714 includes and / or is representative of the platform 716 for resources 718. The platform 716 abstracts underlying functionality of hardware (e.g., servers) and software resources of the cloud 714. The resources 718 include applications and / or data that can be utilized while computer processing is executed on servers that are remote from the computing device 702. Resources 718 also can include services provided over the Internet and / or through a subscriber network, such as a cellular or Wi-Fi network.

[0093] The platform 716 abstracts resources and functions to connect the computing device 702 with other computing devices. The platform 716 also abstracts scaling that facilitates

[0094] In implementations, the platform 716 employs a "machine learning model" configured to implement the techniques described herein. A machine learning model refers to a computer representation that can be adjusted (e.g., trained and retrained) based on inputs to approximate unknown functions. In particular, the term machine learning model can include models that learn from and make predictions on known data by analyzing training data using algorithms to generate outputs that reflect patterns and properties of the training data. Examples of machine learning models include neural networks, convolutional neural networks (CNNs), long short-term memory (LSTM) neural networks, decision trees, and the like.

[0095] Although the application has been described in language specific to structural features and / or methodological acts, it is to be understood that the application defined in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as example forms of implementing the claimed application.

Claims

1. A method comprising: receiving information via input to a user interface, the received information describing an item to be listed for sale; identifying, for each of a plurality of similar items, at least one aspect describing a characteristic of the similar item; generating, using one or more machine learning models, a prompt to initiate a description of the item to be listed for sale with attributes different from the received information based on the identified at least one aspect for each of the plurality of similar items; and presenting the attributes described by the one or more machine learning models for output via the user interface. The information describing the item to be listed for sale includes at least one of an item title or an image of the item to be listed for sale.

2. The method of claim 1, wherein, 3. The method of claim 1, further comprising: generating, using the received information, a latent space embedding of the item to be listed for sale; and identifying the plurality of similar items using the latent space embedding. Identifying, for each of the plurality of similar items, at least one aspect describing a characteristic of the similar item includes comparing each of the plurality of similar items to an item database storing metadata describing listings of items on one or more virtual marketplaces. Generating an item listing for the item to be listed for sale, the item listing including at least some of the attributes described by the one or more machine learning models.

4. The method of claim 3, wherein, At least one of the attributes described by the one or more machine learning models includes a plurality of candidate values, the method further comprising prompting input of feedback via the user interface selecting an appropriate candidate value of the plurality of candidate values for inclusion in the item listing for at least one of the attributes.

5. The method of claim 1, further comprising: The prompt is configured using only text describing the received information and the identified at least one aspect for each of the plurality of similar items.

6. The method of claim 5, wherein, Generating the prompt is performed by filling in a template using the identified at least one aspect for each of the plurality of similar items.

7. The method of claim 1, wherein, The prompt delegates the one or more machine learning models to prioritize the attributes describing the item in order of importance.

8. The method of claim 1, wherein, The prompt delegates the one or more machine learning models to assign a name or value to each of the attributes describing the item to be listed for sale.

9. The method of claim 1, wherein, The prompt causes the one or more machine learning models to describe the item to be listed for sale using at least one attribute not included in the identified at least one aspect for each of the plurality of similar items.

10. The method of claim 1, wherein, 12. A system comprising:

11. The method of claim 1, wherein, one or more processors; and a computer-readable storage medium storing instructions executable by the one or more processors to perform operations comprising: receiving information describing an item to be listed for sale; identifying, for each of a plurality of similar items, at least one aspect describing a characteristic of the similar item; ​ ​ ​ generating, using one or more machine learning models, a prompt to initiate a description of the item to be listed for sale with attributes different from the received information based on the identified at least one aspect for each of the plurality of similar items; and presenting the attributes described by the one or more machine learning models for output via a user interface.

13. The system of claim 12, wherein, The prompt is configured using only text describing the received information and the identified at least one aspect for each of the plurality of similar items.

14. The system of claim 12, wherein, The generating the prompt is performed by filling in a template using the identified at least one aspect for each of the plurality of similar items.

15. The system of claim 12, wherein, The prompt delegates the one or more machine learning models to prioritize the attributes describing the item in order of importance.

16. The system of claim 12, wherein, The prompt delegates the one or more machine learning models to assign a name or value to each of the attributes describing the item to be listed for sale.

17. The system of claim 12, wherein, The prompt causes the one or more machine learning models to describe the item to be listed for sale using at least one attribute not included in the identified at least one aspect for each of the plurality of similar items.

18. The system of claim 12, the operations further comprising: generating an item listing for the item to be listed for sale, the item listing including at least some of the attributes described by the one or more machine learning models.

19. The system of claim 18, wherein, At least one of the attributes described by the one or more machine learning models includes a plurality of candidate values, the operations further comprising prompting for input via the user interface of feedback selecting an appropriate candidate value of the plurality of candidate values for at least one of the attributes to include in the item listing.

20. A computer-readable storage medium storing instructions executable by at least one processor to perform operations comprising: receiving information describing an item to be listed for sale; identifying, for each of a plurality of similar items, at least one aspect describing a characteristic of the similar item; generating, using one or more machine learning models, a prompt to initiate a description of the item to be listed for sale with attributes different from the received information based on the identified at least one aspect for each of the plurality of similar items; generating an item listing for the item to be listed for sale, the item listing including at least some of the attributes described by the one or more machine learning models; and presenting in a user interface the item listing including at least some of the attributes described by the one or more machine learning models.

Images